<p>Phase change materials are exploited in several enabling technologies such as storage class memories, neuromorphic devices and memories embedded in microcontrollers. A key functional property for these applications is the fast crystal nucleation and growth in the supercool liquid phase. Over the last decade, atomistic simulations based on density functional theory (DFT) have provided crucial insights on the early stage of this process. These simulations are, however,<br>restricted to a few hundred atoms for at most a few ns. More recently, the scope of the DFT simulations has been greatly extended by leveraging on machine learning techniques. In our paper, we show that the exploitation of a recently devised neural network potential for the prototypical phase change compound Ge<sub>2</sub> Sb<sub>2</sub>Te<sub>5</sub> (GST) allows simulating the crystallization process in a multimillion atom model at the length and time scales of the real memory device. The simulations provide a vivid atomistic picture of the subtle interplay between crystal nucleation and crystal growth from the crystal/amorphous rim in a model mimicking the operation<br>of the memory in the Wall geometry. Our simulation showes that at realistic conditions of the set operation of the memory in the Wall architecture the crystallization is dominated by growth at the crystal-amorphous interface. Moreover, the simulations have allowed quantifying the distribution of point defects that controls electronic transport, in a very large crystallite grown at the real conditions of the set process of the device.</p>
